conv_transpose1d_op

→TensorImplPtr

int conv_transpose1d_op(const int & x, const int & W, const int & b, int stride_l, int pad_l, int opad_l)

source edit

Python wrappers

pylucid.nn.functional.conv_transpose1dfunction

One-dimensional transposed convolution (fractionally-strided conv).

Upsamples a 1-D batch by inserting stride - 1 implicit zeros between consecutive input positions, then applies a regular cross-correlation against the kernel. This is the primitive used in decoder / generator architectures to lengthen sequences.

Parameters

xTensorImplPtr

Input of shape (B, C_in, L).

WTensorImplPtr

Filters of shape (C_in, C_out, KL).

bTensorImplPtr

Bias of shape (C_out,) or empty.

stride_lint= None

Length-axis upsampling factor. Default: 1.

pad_lint= None

Length-axis padding subtracted from the output extent. Default: 0.

opad_lint= None

Length-axis output padding (one-sided). Must be < stride_l. Default: 0.

Returns

TensorImplPtr

Output of shape (B, C_out, L_out) where $L_\text{out} = (L - 1)\, s - 2 p + KL + p^\text{out}$ .